1
00:00:00,690 --> 00:00:02,100
Hi and welcome back.

2
00:00:02,670 --> 00:00:08,520
In this section, I'll dive into the history of CNN's basically, this is a very high level outline

3
00:00:08,520 --> 00:00:14,490
just showing you a detailed timeline of how CNN's have progressed to evolution.

4
00:00:14,760 --> 00:00:20,640
Just so we know the sequence of when these different networks arose and we can see where we are in our

5
00:00:20,640 --> 00:00:21,190
modern day.

6
00:00:21,210 --> 00:00:26,220
No, I should add this is this has been updated only to 2018.

7
00:00:26,670 --> 00:00:28,740
And the reason for this?

8
00:00:29,040 --> 00:00:35,280
Well, mainly because actually didn't update this, to be fair, but I will be discussing some of the

9
00:00:35,280 --> 00:00:37,200
more modern architectures here.

10
00:00:37,710 --> 00:00:39,560
But I'll go back to this.

11
00:00:39,580 --> 00:00:45,420
The reason I stopped right here, and that's because RTÉ's next, which is a very good CNN network,

12
00:00:45,960 --> 00:00:50,670
it's honestly still has some of the world class performance in image classification.

13
00:00:51,090 --> 00:00:57,780
In fact, there was a 2001 paper that actually showed resonates still performed best, despite all the

14
00:00:57,780 --> 00:01:01,220
fancy architectures that have arose over the years.

15
00:01:01,230 --> 00:01:05,460
So residents are currently my favorite CNN if I had to choose.

16
00:01:06,030 --> 00:01:11,700
But it was the second favorite because it really is just so good and reliable for many different image

17
00:01:11,700 --> 00:01:12,690
recognition tests.

18
00:01:13,230 --> 00:01:15,240
But we'll start at the bottom with this one.

19
00:01:15,240 --> 00:01:19,950
Start with Linnet, which is another classical scene in architecture along with Alex Net.

20
00:01:20,670 --> 00:01:27,180
And then we progressed to video, which came out in 2014, which was the year before I did my Masters

21
00:01:27,180 --> 00:01:27,930
and II.

22
00:01:28,260 --> 00:01:33,120
And I remember these students talking about big like it was so groundbreaking back then.

23
00:01:33,410 --> 00:01:39,120
It doesn't seem that long ago, so you can see how fast things progressed in the deep, deep learning

24
00:01:39,120 --> 00:01:39,420
world.

25
00:01:40,110 --> 00:01:47,130
Next, we had inception and it was we wanted to do tree and current leaders before, but it also resonates.

26
00:01:47,130 --> 00:01:51,930
We'll talk about resonance with groundbreaking resonance, helped residents onto a very complicated

27
00:01:51,930 --> 00:01:55,170
architecture, which we'll discuss, though, and I'll discuss why they work so well.

28
00:01:55,770 --> 00:01:58,050
But they were this of so many problems.

29
00:01:58,050 --> 00:02:04,760
And then basically amidst this type of scene in architecture, so flexible and usable in image recognition,

30
00:02:04,770 --> 00:02:07,200
class and classification and classification.

31
00:02:07,860 --> 00:02:14,460
So also, we take a look at denseness exceptions shuffle that mobile that these were more efficient

32
00:02:14,460 --> 00:02:18,440
models, not groundbreaking state of the art accurate models.

33
00:02:18,450 --> 00:02:19,740
They were just very efficient.

34
00:02:19,740 --> 00:02:27,540
So they were deployed on mobile devices and at a slower CPU type devices where computations computational

35
00:02:27,540 --> 00:02:28,680
fluid mattered a lot.

36
00:02:29,220 --> 00:02:35,160
Then we look at Inception Resonant Inception before RTÉ's next, which I discussed and mobile and that

37
00:02:35,160 --> 00:02:35,790
fit two.

38
00:02:36,240 --> 00:02:40,860
And since then, there's basically been a couple of things that have been happened.

39
00:02:41,520 --> 00:02:46,790
Vision transformers have basically exploded in the last few months in the CNN world.

40
00:02:46,830 --> 00:02:54,080
That's in 2020 to early 2020, a sort of late 2020, early 2021 when it first came on the scene.

41
00:02:54,090 --> 00:02:56,250
And it's still exploding with research right now.

42
00:02:56,670 --> 00:03:00,720
It is getting world class performance and we will discuss efficient transformers here.

43
00:03:01,170 --> 00:03:03,050
I actually should have added it in this diagram.

44
00:03:03,050 --> 00:03:03,900
I'm sorry, I didn't.

45
00:03:04,290 --> 00:03:06,430
But either way, I will discuss it.

46
00:03:06,440 --> 00:03:07,400
It's part of discourse.

47
00:03:07,860 --> 00:03:12,030
Don't you worry, and I'll also say that rest next.

48
00:03:12,120 --> 00:03:18,630
Right before that was still at that point, the last two or three is still getting groundbreaking performance.

49
00:03:18,630 --> 00:03:25,950
So this has a lot that says that maybe we have stagnated on CNN research a bit coming down to the late

50
00:03:26,130 --> 00:03:29,910
2018's early 2020 with no vision.

51
00:03:29,910 --> 00:03:35,910
Transformers have basically progressed to field, and now we've got it getting again.

52
00:03:35,910 --> 00:03:37,580
We're building state of the art schools with them.

53
00:03:38,040 --> 00:03:41,520
So I'll stop talking about this right now and let's get into it.

54
00:03:41,940 --> 00:03:48,990
So let's take a look at Linnet, which is the first classic CNN architecture that was made popular.

55
00:03:49,560 --> 00:03:53,550
So I'll see you in the next section where we take a look at Linet.